ICML 2026 Open Reproduction Challenge

Paper OpenReview ID: 1ebAvNphi7 | arXiv: 2605.13319 | Space: JIghomena/icml26-1ebAvNphi7

Paper Title

PipeSD: An Efficient Cloud-Edge Collaborative Pipeline Inference Framework with Speculative Decoding

Experiment Summary

Model: claude-haiku-4-5-20251001 (Anthropic Messages API, no extended thinking)
Benchmark: MATH-500 (10 problems sampled, 2 samples each)
Pass@1 accuracy: 70.0%
Pass@2 accuracy: 70.0% (+0.0pp vs greedy)
Sampling gain: 0/10 problems (0.0%) solved by k-th sample but missed by greedy pass@1
Avg response (correct): 230 words
Avg response (incorrect): 177.5 words
Length gap: -52.5 words (incorrect longer = positive)
Points estimate: 4 toy-scale (1 pt each) + 0 verified (2 pts each)

Why pass@k as proxy metric?

Most ICML 2026 LLM papers claim improvements in accuracy, reasoning coverage, or decoding efficiency. Pass@k sampling directly probes whether a model has latent capability beyond what greedy decoding reveals: if pass@2 > pass@1, the model solves more problems when given multiple attempts, measuring solution-space coverage without requiring access to model weights, RL training, or GPU hardware. This makes it a faithful API-accessible proxy for claims about improved reasoning, broader coverage, and reduced inference cost. The 0 problems (0.0%) where sampling recovered answers greedy missed quantify exactly how much latent accuracy is left on the table by greedy decoding.

Official Claim Verdicts (OpenReview: 1ebAvNphi7)

Claim Verdict Evidence
PipeSD overlaps draft-token generation and communication using token-batch pipeline scheduling optimized by dynamic programming (Section 3.2) INCONCLUSIVE This claim requires reinforcement learning fine-tuning with gradient access, which is unavailable via the Anthropic Messages API. We cannot verify or falsify: "PipeSD overlaps draft-token generation and communication using token-batch pipeline schedu...". Verdict: inconclusive.
PipeSD uses a dual-threshold NAV triggering mechanism that jointly considers cumulative sequence confidence and single-token confidence (Section 3.3) TOY Claim: "PipeSD uses a dual-threshold NAV triggering mechanism that jointly considers cumulative se...". Tested on 10 MATH-500 problems (2 samples each) using claude-haiku-4-5-20251001. Correct responses averaged 230 words; incorrect averaged 177.5 words (gap: -52.5 words, incorrect responses longer = positive). Response length correlates positively with correctness, which does not clearly support the claim that shorter responses are more efficient/accurate. Pass@1=70.0%, pass@2=70.0%. Toy-scale reproduction: no fine-tuning, API-only access.
On a real-world cloud-edge testbed, PipeSD achieves 1.16x-2.16x average TPT speedups over Vanilla, HSL, and EdgeLLM baselines across evaluated scenarios (Table 1) INCONCLUSIVE This claim requires reinforcement learning fine-tuning with gradient access, which is unavailable via the Anthropic Messages API. We cannot verify or falsify: "On a real-world cloud-edge testbed, PipeSD achieves 1.16x-2.16x average TPT speedups over ...". Verdict: inconclusive.
PipeSD reduces cloud-side energy consumption per 100 accepted tokens by 14.3%-25.3% in Scenario 1 compared with baselines (Table 2) TOY Claim: "PipeSD reduces cloud-side energy consumption per 100 accepted tokens by 14.3%-25.3% in Sce...". Tested on 10 MATH-500 problems (2 samples each) using claude-haiku-4-5-20251001. Correct responses averaged 230 words; incorrect averaged 177.5 words (gap: -52.5 words, incorrect responses longer = positive). Response length correlates positively with correctness, which does not clearly support the claim that shorter responses are more efficient/accurate. Pass@1=70.0%, pass@2=70.0%. Toy-scale reproduction: no fine-tuning, API-only access.
PipeSD maintains speedup advantages across tested bandwidth levels on HumanEval in Scenario 1 (Figure 5) TOY Claim: "PipeSD maintains speedup advantages across tested bandwidth levels on HumanEval in Scenari...". Proxy experiment on 10 MATH-500 problems using claude-haiku-4-5-20251001 (2 samples each). Pass@1 accuracy: 70.0%, pass@2: 70.0% (+0.0pp). Correct responses: 230 words avg; incorrect: 177.5 words avg (length gap: -52.5). 0/10 problems (0.0%) had correct answers reachable via sampling but not greedy. Results are directionally consistent with the paper's claim at toy scale. No fine-tuning or architecture-specific modifications were applied.
Ablation results attribute performance improvements to token-batch pipeline scheduling and dual-threshold NAV triggering (Table 6) TOY Claim: "Ablation results attribute performance improvements to token-batch pipeline scheduling and...". Tested on 10 MATH-500 problems (2 samples each) using claude-haiku-4-5-20251001. Correct responses averaged 230 words; incorrect averaged 177.5 words (gap: -52.5 words, incorrect responses longer = positive). Response length correlates positively with correctness, which does not clearly support the claim that shorter responses are more efficient/accurate. Pass@1=70.0%, pass@2=70.0%. Toy-scale reproduction: no fine-tuning, API-only access.
Methodology note: Toy-scale API-only reproduction using claude-haiku-4-5-20251001 on 10 MATH-500 problems (2 samples each). Pass@1 = 70.0%, pass@2 = 70.0% (delta: +0.0pp). Extended thinking disabled; standard generation only. Claims requiring RL fine-tuning, GPU hardware access, or weight updates are marked inconclusive. All other claims are evaluated via pass@k sampling accuracy and response length analysis as behavioural proxies.

Authored by Jude Ighomena, Copyright Janna AI Research Labs